Relativistic GAN using Receptive Field Block for Single Image Super-Resolution with improved Perceptual Quality
Authors
P. S. HrishikeshDensen PuthusseryAkhil K. A.
Abstract
Generative adversarial networks (GAN) are proved to be extremely useful to solve the Single image super resolution(SISR) problem as they can recover the finer texture details even with large upsamplng factors. In this paper, we propose a deep network architecture using a relativistic generative adversarial network (V-SRGAN) with receptive field block (RFB) for image super-resolution having good perceptual quality. Our generator network uses multi-scale RFBs which are capable of extracting the coarse and finer features from the input low resolution image to recover the super resolved image with finer details and textures. It is initially trained on mean absolute error (MAE) followed with relativistic average GAN (RaGAN) loss for both discriminator and generator. Training based on RaGAN loss enables the network to map the low resolution images to more realistic high-resolution counterparts. The proposed network was able to attain better results in terms of PSNR and learned perceptual image patch similarity (LPIPS) metric in comparison with the other GAN based methods.
Why we worked on this
Super-resolution is where the gap between a good metric and a good image is most obvious. A network can improve PSNR while producing something that looks worse to a person, which is why this work is evaluated on perceptual similarity as well. Deciding what “better” means before optimising for it is not an academic nicety; it is the difference between a model that satisfies a benchmark and one that satisfies a user.
Where to start
The AI Modernization Assessment is two weeks and $5,000. Not a code audit. You get a prioritized roadmap, effort and cost ranges, and a written list of the workflows where AI would add risk without adding value.
Book the assessment